Unsupervised degradation learning model for endoscopic video super-resolution

Abstract The lack of paired high-resolution (HR) and low-resolution (LR) data poses a significant challenge for endoscopic video super-resolution. Conventional methods typically rely on predefined degradation functions (e.g., bicubic downsampling) to construct paired training data, which fail to accurately model complex degradations in real-world scenarios, resulting in a distribution gap between synthetic and real LR data, which in turn compromises the generalization capability of super-resolution models in practical applications. Although some approaches attempt to learn downsampling models to generate pseudo-paired data, in endoscopic scenarios, the degradation process is often complex and diverse, making the generated results prone to artifacts and color distortions. In addition, these methods usually process video frames independently, making it difficult to preserve temporal coherence during the degradation process. To address these issues, we propose an unsupervised degradation learning method. By leveraging unpaired clinical HR and LR video frames, the proposed method learns a degradation mapping from the HR domain to the real LR domain, thereby generating data consistent with the real LR distribution. We design a cyclic degradation learning framework to model the mapping between the HR and LR domains. By introducing a low-frequency loss and an adaptive data loss, the degradation simulation is guided to preserve structural content information while suppressing high-frequency artifacts. For video sequences, we introduce a Temporal Loss to promote motion consistency between the generated LR frames and their corresponding HR frames, which helps maintain spatiotemporal consistency during the degradation process. Experimental results demonstrate that, compared with conventional degradation modeling methods, the proposed downsampling model improves the reconstruction performance of existing video super-resolution methods on synthetic datasets and shows certain generalization potential on real clinical endoscopic videos.